Insight

Write once, patch everywhere

18 November 2026

Your Java patch will reassure you, right up until you find the JDK baked into an image nobody has rebuilt since spring.

Imagine the car maker recalls a model, but only the ones fitted with a particular engine. Very helpful. Except nobody wrote down which engine went in which car, half the fleet is leased, and the hire company can swap you into a different car next week.

That’s Java4Shell. It’s a memory-corruption flaw in the OpenJDK Base64 decoder, in builds from 17.0.6 onwards, including 21, 25 and 27. A short, malformed input makes hand-tuned code write 62 bytes past the end of a buffer before Java gets round to rejecting it. Base64 often gets decoded before anyone logs in, so the researchers at Calif, who found it, showed code execution before login on default installs of Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate. Your app’s code can be spotless and still be exposed, because the problem is in the runtime underneath it. Some people are calling it Java2Shell. Same bug.

Does it matter what chip it runs on?

This is the bit that makes Java4Shell different. The flaw lives in an optimisation that only switches on for x86-64 processors with AVX-512 VBMI: Intel Xeon Ice Lake and newer, and AMD Zen 4 and newer. JFrog puts it plainly: “ARM64 machines (AWS Graviton, Apple Silicon) are not affected.” On AWS, that means the m6i, c6i, c7i, m7a and c7a families are in, and the Graviton fleet is out.

So the same container image can be exposed on one instance type and clear on another. Your vulnerability scanner sees the image. It doesn’t see the chip. (Spoiler: neither does your CMDB!)

Here’s where I’d be careful, though. A clear chip today isn’t a fix. Someone resizes an instance, a node group picks a new family, Fargate lands your task on different hardware, and the exposure is back with nothing in the change log to say so. Qualys says it best: “Use the processor condition to order the work, not to close findings.”

So why isn’t patching enough?

Because Java doesn’t live in one place. dnf upgrade fixes the Java your operating system installed. It doesn’t touch the JDK your identity provider ships in its own folder, the one baked into a container image when it was built, or the one inside a vendor’s Marketplace appliance. Those stay on the old build until someone rebuilds the image or the vendor ships an update.

And even a fixed host isn’t fixed until the process restarts. A JVM that’s been up since October is still running the October build, whatever the package manager says. Check versions, not vibes.

The researchers say they found the bug with AI help in about an hour, and had a first working exploit two hours later, for under $200. The time between “published” and “someone’s trying it on your login page” isn’t going to be measured in weeks.

What do I do this week?

  1. Check the chip. On a Linux host, an affected processor shows all four of these flags:

    grep -o -w -E 'avx512f|avx512bw|avx512vbmi|bmi2' /proc/cpuinfo | sort -u

    On AWS you can do it from the API instead. List instance types, and check the architecture of your Java Lambda functions (arm64 is clear):

    aws lambda list-functions --query "Functions[?starts_with(Runtime || '', 'java')].[FunctionName,Runtime,Architectures[0]]" --output table
  2. Find every copy of Java, not just the one on the PATH. Look for every java binary on your hosts, and look inside your images:

    find / -xdev -type f -name java -perm -u+x 2>/dev/null
    syft <image> | grep -iE 'jdk|jre'
  3. Check the build that’s running, not the one installed. jcmd <pid> VM.version tells you what a live JVM is actually running. You want Oracle 17.0.21.1, 21.0.13.1, 25.0.5.1 or 27.0.1.1, or Corretto 17.0.21.1.2, 21.0.13.1.2, 25.0.5.1.2 or 27.0.1.1.4. JFrog notes there are no fixes for branches 18 to 20, 22 to 24 and 26, so those need moving to a supported line.

  4. Rank by who can reach it. Identity services first, then CI/CD and collaboration tools, then anything else that handles logins or tokens, then internal-only services. Affected chips go to the top of each group.

  5. Mitigate what has to wait. Start the JVM with -XX:-UseBASE64Intrinsics, which falls back to the standard decoder. Test first, because Base64-heavy code may run slower, apply it on affected x86 hosts only, restart, and check it’s set with jcmd <pid> VM.flags. Write down who owns taking it off once the real fix lands.

  6. Rebuild, don’t patch the container. Bump the base image or the JDK in your Dockerfile and redeploy. Patch a running container and it’ll come straight back on the next deploy.

Remember, step two decides how long everything else takes. And on Amazon Linux, check AWS’s own advisory page for package status rather than trusting a headline.

Can you really triage that much at once?

You can. For one digital marketing and directory services business, our engine triaged and resolved around 10,000 AWS security findings in a single run, with no added headcount and higher-risk changes held for human approval. Ranking every Java runtime by chip, build and reachability is the same kind of job: tedious for people, quick for a machine that already has the map.

Anyway, don’t be like me and assume the Graviton migration you did last year covered everything. Here at Firemind we’ve built the IT Operations Engine to keep a current map of what runs across your cloud and your code, fix it in order inside your rules, and hold higher-risk changes for a human to approve, with every step recorded. Start with a free, read-only assessment: we’ll find every Java runtime in one AWS account, check the chip and the build actually running, and rank the lot.

Find every Java runtime in your AWS estate, free → AWS Marketplace

Sources

  • Qualys Threat Research Unit, Saeed Abbasi, 17–18 November 2026: the flaw in the AVX-512 Base64 decoder and the 62-byte overwrite; affected builds from 17.0.6, including 21, 25 and 27; the AVX-512 VBMI processor condition and the four CPU flags; Oracle’s fixed builds and Corretto 17.0.21.1.2; the -XX:-UseBASE64Intrinsics mitigation and its performance caveat; restarting and checking the running process; bundled JDKs; the products Calif demonstrated; Calif’s account of finding the bug with AI help in about an hour, an exploit two hours later and under $200 in costs; the advice to use the processor to order work, not close findings.
  • JFrog Security Research, 17 November 2026: both names in use; fixed versions per branch and no fixes for branches 18 to 20, 22 to 24 and 26; standard Java 8 and 11 not affected; ARM64, including AWS Graviton, not affected; the AWS instance families that meet the condition; the advice to check JDKs bundled in products and container images; the advice to apply mitigations on x86 only.
  • Amazon Linux Security Center, CVE-2026-114938, public 17 November 2026: OpenJDK 17.0.6 and later on x86-64 with specific CPU capabilities; ARM64 not affected; Corretto package status by Amazon Linux version.
  • The Hacker News, 18 November 2026: Corretto 17.0.21.1.2, 21.0.13.1.2, 25.0.5.1.2 and 27.0.1.1.4 carry the fix, per the Corretto changelogs.
  • Firemind case study: around 10,000 AWS security findings triaged and resolved in a single engine run, with higher-risk changes held for human approval.

Frequently asked questions.

Java4Shell, also reported as Java2Shell, is a memory-corruption flaw in the OpenJDK Base64 decoder, published on 17 November 2026. A short, malformed Base64 input makes optimised code write past the end of a buffer. Researchers at Calif showed code execution before login on default installs of products including Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate.

OpenJDK-based builds from 17.0.6 onwards, including 21, 25 and 27, from every major vendor, but only on x86-64 processors with AVX-512 VBMI: Intel Xeon Ice Lake and newer, and AMD Zen 4 and newer. ARM64, including AWS Graviton, isn't affected. Standard Java 8 and 11 aren't affected.

Oracle's fixed builds are 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1. Amazon Corretto's are 17.0.21.1.2, 21.0.13.1.2, 25.0.5.1.2 and 27.0.1.1.4. JFrog notes there are no fixes for branches 18 to 20, 22 to 24 and 26, so those need an upgrade to a supported line or the interim mitigation.

Start the JVM with -XX:-UseBASE64Intrinsics, which makes Java fall back to its standard Base64 decoder. Test it first, because code that decodes a lot of Base64 may run slower, then restart the process and check the flag is set on the running JVM. Apply it on affected x86 hosts only.

View all insights

CONTACT US

Start with a focused conversation about your environment.

We help you build, optimise and run AI that delivers measurable results.

Your benefits:

  • Outcome-driven - Measurable business impact
  • Expert-led - Hands-on delivery from senior practitioners
  • Secure by design - Your data and compliance requirements first
  • Fast to value - From discovery to production in weeks

What happens next?

Let's talk

A 20-minute focused session on your goals and current situation.

We propose

A clear plan and scope tailored to your priorities.

You decide

No obligation - move forward when the time is right.

No obligation - just a focused 30-minute discussion about your goals.

We'll only use your details to respond to your enquiry. No newsletters unless you ask for them.